nanodet

Why we included this project

NanoDet is a strong choice when object detection needs to run in real time on a phone or edge device without a GPU. It's a full training and inference toolkit for a tiny anchor-free detector; the smallest quantized model is under a megabyte and still hits about 97 fps on a mobile ARM CPU. The repo includes PyTorch training code and configs, pretrained COCO checkpoints, a model zoo, and working demos for ncnn, MNN, and OpenVINO, plus an Android sample app built on ncnn. That makes it practical for on-device apps, surveillance cameras, and robotics, where you're always trading accuracy against latency and memory. Benchmark tables compare honestly against YOLO-tiny variants, and the docs walk through install, training, and deployment without much friction.